How to stop LLM (OpenAI) assistant to tell me that I can ask for more information EVERY SINGLE TIME.

[plain shoal]

Hi there 😉

I already tried to stop it by adding a section to the prompt.
That it shouldn't be too wordy at the end of the reply and doesn't need to tell me all the time that I can ask for more information if needed (or something similar).
I already know it, at least since the 94835756. time I heard it. 😵‍💫

Is this something that happens to others too, or is this something in my prompt that leads to this behavior?

If you want to talk via voice to the assist and ask a few follow up quesitons this is way too time consuming. I already start to take my phone out of the pocket and look it up myself. 😄

[frigid minnow]

What is your prompt?

And do you use Ollama or that home-llm integration?

This is my prompt with Qwen_3_8b_Q8. It rarely does this behavior you describe, instead only pinging me when needed

[plain shoal]

I use OpenAI 4o-mini

But your question about the prompt is a good hint.
Will create a new assistant with the default prompt and compare the behavior.

If it's the same it might be related to HAs default prompt around our custom prompt (maybe related to the ongoing conversation thing, where assist can ask you questions to keep the mic input active).

Will report back after that test ...

[lone hinge]

My LLM assistant based on qwen 3 14b with a default HA prompt is also annoying with frequent useless “let me know how I can help” and emojis which are then spoken out loud by the Voice PE, despite “respond only in plain text” in my prompt. It’s almost like this thing has a mind of its own, right?

[lone hinge]

Your prompt is massive, was there some guide you followed to create it? I thought the instructions how to handle home assistant are baked into the Ollama integration and/or trained into the model itself. I’ll try playing with your prompt also, see if it helps make it better or worse.

[plain shoal]

I don’t have the default tools described here, but I mention some of my custom tools that it should use instead of looking things up itself. So it really is aware of these use cases.

Also a lot stuff that it has to know about the house, special devices or the family.

[frigid minnow]

I found it on some github that I can't find again.

But it's massively increased the success rate (according to my gut feeling)

especially the plex script lines

it calls that script properly almost every time